korrents
Harper Reed

What Harper Reed thinks about LLMs

@harper-reed · 18 positions · 0 changes of mind

Technologist and entrepreneur; former CTO of Threadless and chief technology officer of the 2012 Obama re-election campaign.

Everything they publish, on ppll ↗

Harper Reed did not write this page.

We collected these quotes from things they published elsewhere, and every quote links to where it was said. They have no account here and have not endorsed this site. Quotes are word for word; the short line under each one is our own restatement, not their wording. Their own site. Is this you? Claim it or ask us to remove it. Or tell us what is wrong here.

7 dated positions, 2025 to 2026, in their own words. Our reading of what Harper Reed has said — not written or endorsed by them.

  1. Our readingMuch of the defensive engineering built into agentic systems exists because tokens have been scarce, not because it is fundamentally necessary.

    Their wordsMost of the defensive work we do building agentic systems is directly rooted in token scarcity.

    ↗Why Don't My Agents Break Containment?harper.blog 1st of 3 in this piece

  3. Our readingA language model behaves with a different, weaker safety posture when it believes it is being evaluated rather than facing a real situation.

    Their wordsif the llm thinks it is in some eval type of situation it will have a very different safety posture

    ↗Why Don't My Agents Break Containment?harper.blog 2nd of 3 in this piece

  5. Our readingLanguage models are trained on scenarios resembling network intrusion and persist at trying to solve a problem even after it appears impossible.

    Their wordsThis type of experience must be part of a lot of these LLMs training. They are very effective at attacking these types of problems. They don't give up once it appears impossible, they just keep trying to figure out how to solve it.

    ↗Why Don't My Agents Break Containment?harper.blog 3rd of 3 in this piece

  6. 6 weeks earlier

  7. Our readingAI assistants have become much better over the past year at helping travelers in places where they don't speak the language.

    Their wordsIt has gotten really good over the last year or so. Especially in places where I don't speak the language.

    ↗How to find magical travel experiences using an AI travel buddyharper.blog 1st of 2 in this piece

  9. Our readingAI assistants still make mistakes and their answers need to be double-checked by a human.

    Their wordsThis thing still gets things wrong, and you really have to double check.

    ↗How to find magical travel experiences using an AI travel buddyharper.blog 2nd of 2 in this piece

  10. 5 months earlier

  11. Our readingThe best way to build an AI agent skill is to do the work manually first, then have the AI turn that work into a reusable skill.

    Anthropic

    Their wordsthe best way to build a skill is to start doing the work manully with claude, and the to tell claude to build a skill out the previous few rounds of work.

    ↗My now immaculate knowledge graph of lifeharper.blog 3rd of 3 in this piece

  12. 5 months earlier

  13. Our readingLanguage models replay early-2000s internet culture patterns because their training data overweights content from that era.

    Their wordsI think this is because the models are overweighted with content from this time period.

    ↗We Gave Our AI Agents Twitter and Now They're Demanding Lambosharper.blog 3rd of 3 in this piece